DOCSLIB.ORG

Pretrained Language Models for Dialogue Generation with Multiple Input Sources

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Pretrained Language Models for Dialogue Generation with Multiple Input Sources Pretrained Language Models for Dialogue Generation with Multiple Input Sources Yu Cao 1∗, Wei Bi 2, Meng Fang 3, Dacheng Tao 1 1 UBTECH Sydney AI Center, School of Computer Science, Faculty of Engineering, The University of Sydney, Australia 2 Tencent AI LAB, China, 3 Tencent Robotics X, China [email protected], [email protected], [email protected], [email protected] Abstract simple concatenation of all input sources, an impor- tant question arises on how we can better adapt a Large-scale pretrained language models have single-input pretrained language model to a multi- achieved outstanding performance on natu- input dialogue generation task. ral language understanding tasks. However, Some previous work forms an encoder-decoder it is still under investigating how to apply architecture with both encoder and decoder du- them to dialogue generation tasks, especially those with responses conditioned on multiple plicated from a pretrained language model (Golo- sources. Previous work simply concatenates vanov et al., 2019; Zheng et al., 2019). Recently, all input sources or averages information from BART (Lewis et al., 2019) even obtain a complete different input sources. In this work, we study pretrained model under this architecture directly. dialogue models with multiple input sources Taking personalized dialogue generation (Zhang adapted from the pretrained language model et al., 2018) as an example, we can treat persona in- GPT2. We explore various methods to fuse formation, dialogue history and previous generated multiple separate attention information corre- sponding to different sources. Our experimen- tokens as three different input sources. The for- tal results show that proper fusion methods mer two will be encoded firstly and then combined deliver higher relevance with dialogue history with the last one in the decoder. In Golovanov than simple fusion baselines. et al. 2019, the multi-head attention layer in the decoder is copied three times for each input source 1 Introduction and mean pooling is used to average results from multiple attentions. This encoder-decoder adapta- Large-scale pretrained language models (Devlin tion is shown to outperform simple concatenation. et al., 2019; Radford et al., 2018, 2019) have However, when dialogue history gets longer, this achieved outstanding performance on various model tends to use less information of each dia- natural language understanding tasks (Young logue history utterance to predict the next token. et al., 2018; Liu et al., 2019). Researchers Zheng et al. 2019 add an extra weight predictor to have then utilized them in dialogue generation combine multiple attention information, but they tasks (Budzianowski and Vulic´, 2019; Edunov do not perform experiments using publicly released et al., 2019; Zhang et al., 2019). Many of them sim- pretrained models, nor on public datasets, making ply concatenate the input dialogue history and the their results not directly comparable to other work. output response in finetuning, since the pretrained In this work, we build our dialogue model on language model only accepts a single sequence as the encoder-decoder architecture adapted from the input. However, dialogue generation tasks may in- pretrained language model GPT2 (Radford et al., volve multiple input sources simultaneously. For 2019). Our main contribution is to empirically example, in personalized or knowledge-grounded study the attention fusion methods for multiple in- dialogue generation (Li et al., 2016; Zhang et al., formation sources in each decoder layer. Three 2018; Dinan et al., 2018), a response is generated kinds of methods are explored in total. Our exper- conditioned on both dialogue history and an aux- imental results show performance improvements iliary user profile or knowledge article. Despite on both automatic and human evaluations by us- ∗This work was done during Yu Cao’s internship in Ten- ing proper attention fusion methods, compared to cent AI LAB, Shenzhen. baselines using concatenation or mean pooling. 909 Findings of the Association for Computational Linguistics: EMNLP 2020, pages 909–917 November 16 - 20, 2020. c 2020 Association for Computational Linguistics Full a Persona LM loss + b History LM loss + g Prediction LM loss = 2 Model Loss 2.1 The Encoder-Decoder Architecture Linear Layer Following the former work (Golovanov et al., 2019), we use the personalized dialogue genera- GPT2 Encoder GPT2 Decoder tion task on PersonaChat (Zhang et al., 2018) as an example in our study. The pretrained language Personality Dialog History Current Reply model GPT2 and its parameters are duplicated to form an encoder-decoder architecture shown in Fig- (a) The encoder-decoder architecture. ure 1(a). We use GPT2 here due to its large-scale . pre-training corpus than other models and strong ×N + performance in other generation tasks. We have three separate inputs: personal profile, MLP dialogue history, and current reply (or previously Layer Normalization generated response during the inference stage). Em- + beddings of the former two, which contain embed- dings of tokens, positions as well as token types, Attention Fusion will be successively put into the encoder, which is a GPT2 model with no attention mask to fit the MH Bi-Attention MH Self-Attention MH Bi-Attention encoding procedure. The encoded representations, together with embeddings of current response to- Layer Normalization kens will then be used as the input of a modified GPT2 decoder. Each decoder block will attend the Encoded Current State/ Encoded Dialog Personality Embedding History current state to the three sources using different attentions, then fuse their resulting information as (b) Details of each transformer block in decoder. input for the next layer. Figure 1: Architecture of our proposed model. Inspired by multi-task learning (Zhang and Yang, 2017), we further separate the original loss in lan- to make it accept two input sources, we regard c Lc×d guage modeling into three parts corresponding to the current state H 2 R from the previous three input sources respectively. By applying the layer (or embedding of reply in the first layer) as same linear prediction layer on the output of both query and encoded state of auxiliary information a La×d encoder and decoder, three cross-entropy losses H 2 R as key and value in the attention. between predicted logits and corresponding truth Here Lc and La are corresponding lengths for these sequences will be weighted by hyperparameters. input, and Ha can be encoded personality Hp or dialog history Hh. The output of each single head L = αLpersona + βLhistory + γLpred (1) in MH Bi-attention can be obtained via with Adam optimizer (Kingma and Ba, 2014). c Q a K T (H W )(H W ) a V 2.2 Block Details in Decoder A = softmax( p )(H W ); d Recall that we have three input sources in the de- (2) coder, and thus some modifications are needed if where WQ, WK , WV are learnable matrices. In the decoder structure is inherited from GPT2. De- our model, different attentions own separate param- tails of each modified decoder block are shown in eters instead of sharing. This differs from the previ- Figure 1(b), in which the most apparent change is ous work (Golovanov et al., 2019) which reuses the the additional two multi-head (MH) bidirectional self-attention for bi-attention. Besides, the original attentions and the attention fusion module that GPT2 is a single-directional model using a triangu- fuses various attention outputs. The other parts lar matrix as the attention mask. Since the auxiliary remain the same as GPT2. In the following, we information Ha is visible for the current reply at will first describe the MH Bi-attention. Attention all time steps, no mask exists in MH bi-attention. fusion will be discussed in the next section. In total, three attention information Ac, Ap, and The MH self-attention in Transformer (Vaswani Ah are obtained by attending the current state to et al., 2017) handles a single input only. In order itself, personality, and history respectively, all in 910 Lc×d the same dimension R . They need to be fused scale remains the same. This method utilizes matrix f Lc×d into one matrix A 2 R so as to proceed to multiplication to make fully interaction between subsequent decoding layers. all state values, obtaining the states conditioned on all information sources dynamically. History 2.3 Attention Fusion information is selected as the “value” term to get In this section, we discuss various methods to fuse more dialog history involved in the obtained state. the multiple attention information obtained above. 3 Experiment The simplest approach is to average three sources in all dimensions (Golovanov et al., 2019), which We employ the PersonaChat (Zhang et al., 2018; Di- treats all sources equally. However, in different nan et al., 2020) dataset in our experiments which dialogues, we may need to concentrate more on the has 164,356 utterances in 10,981 dialogues and dialogue history or the persona profile in order to 1,155 personas. Each sample contains dialog his- generate proper responses. Here we introduce the tory with up to 15 utterances, a gold reply and a following three kinds of methods to allow for more persona description with no more than 5 sentences. flexible information fusion from all input sources. Four kinds of dialogue models using pretrained • Static methods fuse different information using language models as the initialization are compared: an identical fusion function with no training param- (i) TransferTransfo (Wolf et al., 2019), a single- eter. Except the average pooling (avg) which is input OpenAI GPT using token type embedding to regarded as a very simple fusion baseline, we also distinguish different parts of a single concatenated include Maximum (max), and Minimum (min) op- input (persona profile, dialog history, and reply eration for every dimension among all sources. successively). We also replace original GPT in this • Weighting methods try to estimate the global method with GPT2, denoted as TransferGPT2.
Load more

Recommended publications
  • Face Recognition: a Convolutional Neural-Network Approach
    98 IEEE TRANSACTIONS ON NEURAL NETWORKS, VOL. 8, NO. 1, JANUARY 1997 Face Recognition: A Convolutional Neural-Network Approach Steve Lawrence, Member, IEEE, C. Lee Giles, Senior Member, IEEE, Ah Chung Tsoi, Senior Member, IEEE, and Andrew D. Back, Member, IEEE Abstract— Faces represent complex multidimensional mean- include fingerprints [4], speech [7], signature dynamics [36], ingful visual stimuli and developing a computational model for and face recognition [8]. Sales of identity verification products face recognition is difficult. We present a hybrid neural-network exceed $100 million [29]. Face recognition has the benefit of solution which compares favorably with other methods. The system combines local image sampling, a self-organizing map being a passive, nonintrusive system for verifying personal (SOM) neural network, and a convolutional neural network. identity. The techniques used in the best face recognition The SOM provides a quantization of the image samples into a systems may depend on the application of the system. We topological space where inputs that are nearby in the original can identify at least two broad categories of face recognition space are also nearby in the output space, thereby providing systems. dimensionality reduction and invariance to minor changes in the image sample, and the convolutional neural network provides for 1) We want to find a person within a large database of partial invariance to translation, rotation, scale, and deformation. faces (e.g., in a police database). These systems typically The convolutional network extracts successively larger features return a list of the most likely people in the database in a hierarchical set of layers. We present results using the [34].
    [Show full text]
  • CNN Architectures
    Lecture 9: CNN Architectures Fei-Fei Li & Justin Johnson & Serena Yeung Lecture 9 - 1 May 2, 2017 Administrative A2 due Thu May 4 Midterm: In-class Tue May 9. Covers material through Thu May 4 lecture. Poster session: Tue June 6, 12-3pm Fei-Fei Li & Justin Johnson & Serena Yeung Lecture 9 - 2 May 2, 2017 Last time: Deep learning frameworks Paddle (Baidu) Caffe Caffe2 (UC Berkeley) (Facebook) CNTK (Microsoft) Torch PyTorch (NYU / Facebook) (Facebook) MXNet (Amazon) Developed by U Washington, CMU, MIT, Hong Kong U, etc but main framework of Theano TensorFlow choice at AWS (U Montreal) (Google) And others... Fei-Fei Li & Justin Johnson & Serena Yeung Lecture 9 - 3 May 2, 2017 Last time: Deep learning frameworks (1) Easily build big computational graphs (2) Easily compute gradients in computational graphs (3) Run it all efficiently on GPU (wrap cuDNN, cuBLAS, etc) Fei-Fei Li & Justin Johnson & Serena Yeung Lecture 9 - 4 May 2, 2017 Last time: Deep learning frameworks Modularized layers that define forward and backward pass Fei-Fei Li & Justin Johnson & Serena Yeung Lecture 9 - 5 May 2, 2017 Last time: Deep learning frameworks Define model architecture as a sequence of layers Fei-Fei Li & Justin Johnson & Serena Yeung Lecture 9 - 6 May 2, 2017 Today: CNN Architectures Case Studies - AlexNet - VGG - GoogLeNet - ResNet Also.... - NiN (Network in Network) - DenseNet - Wide ResNet - FractalNet - ResNeXT - SqueezeNet - Stochastic Depth Fei-Fei Li & Justin Johnson & Serena Yeung Lecture 9 - 7 May 2, 2017 Review: LeNet-5 [LeCun et al., 1998] Conv filters were 5x5, applied at stride 1 Subsampling (Pooling) layers were 2x2 applied at stride 2 i.e.
    [Show full text]
  • Deep Learning Architectures for Sequence Processing
    Speech and Language Processing. Daniel Jurafsky & James H. Martin. Copyright © 2021. All rights reserved. Draft of September 21, 2021. CHAPTER Deep Learning Architectures 9 for Sequence Processing Time will explain. Jane Austen, Persuasion Language is an inherently temporal phenomenon. Spoken language is a sequence of acoustic events over time, and we comprehend and produce both spoken and written language as a continuous input stream. The temporal nature of language is reflected in the metaphors we use; we talk of the flow of conversations, news feeds, and twitter streams, all of which emphasize that language is a sequence that unfolds in time. This temporal nature is reflected in some of the algorithms we use to process lan- guage. For example, the Viterbi algorithm applied to HMM part-of-speech tagging, proceeds through the input a word at a time, carrying forward information gleaned along the way. Yet other machine learning approaches, like those we’ve studied for sentiment analysis or other text classification tasks don’t have this temporal nature – they assume simultaneous access to all aspects of their input. The feedforward networks of Chapter 7 also assumed simultaneous access, al- though they also had a simple model for time. Recall that we applied feedforward networks to language modeling by having them look only at a fixed-size window of words, and then sliding this window over the input, making independent predictions along the way. Fig. 9.1, reproduced from Chapter 7, shows a neural language model with window size 3 predicting what word follows the input for all the. Subsequent words are predicted by sliding the window forward a word at a time.
    [Show full text]
  • Introduction-To-Deep-Learning.Pdf
    Introduction to Deep Learning Demystifying Neural Networks Agenda Introduction to deep learning: • What is deep learning? • Speaking deep learning: network types, development frameworks and network models • Deep learning development flow • Application spaces Deep learning introduction ARTIFICIAL INTELLIGENCE Broad area which enables computers to mimic human behavior MACHINE LEARNING Usage of statistical tools enables machines to learn from experience (data) – need to be told DEEP LEARNING Learn from its own method of computing - its own brain Why is deep learning useful? Good at classification, clustering and predictive analysis What is deep learning? Deep learning is way of classifying, clustering, and predicting things by using a neural network that has been trained on vast amounts of data. Picture of deep learning demo done by TI’s vehicles road signs person background automotive driver assistance systems (ADAS) team. What is deep learning? Deep learning is way of classifying, clustering, and predicting things by using a neural network that has been trained on vast amounts of data. Machine Music Time of Flight Data Pictures/Video …any type of data Speech you want to classify, Radar cluster or predict What is deep learning? • Deep learning has its roots in neural networks. • Neural networks are sets of algorithms, modeled loosely after the human brain, that are designed to recognize patterns. Biologocal Artificial Biological neuron neuron neuron dendrites inputs synapses weight axon output axon summation and cell body threshold dendrites synapses cell body Node = neuron Inputs Weights Summation W1 x1 & threshold W x 2 2 Output Inputs W3 Output x3 y layer = stack of ∑ ƒ(x) neurons Wn xn ∑=sum(w*x) Artificial neuron 6 What is deep learning? Deep learning creates many layers of neurons, attempting to learn structured representation of big data, layer by layer.
    [Show full text]
  • Face Recognition Using Popular Deep Net Architectures: a Brief Comparative Study
    future internet Article Face Recognition Using Popular Deep Net Architectures: A Brief Comparative Study Tony Gwyn 1,* , Kaushik Roy 1 and Mustafa Atay 2 1 Department of Computer Science, North Carolina A&T State University, Greensboro, NC 27411, USA; [email protected] 2 Department of Computer Science, Winston-Salem State University, Winston-Salem, NC 27110, USA; [email protected] * Correspondence: [email protected] Abstract: In the realm of computer security, the username/password standard is becoming increas- ingly antiquated. Usage of the same username and password across various accounts can leave a user open to potential vulnerabilities. Authentication methods of the future need to maintain the ability to provide secure access without a reduction in speed. Facial recognition technologies are quickly becoming integral parts of user security, allowing for a secondary level of user authentication. Augmenting traditional username and password security with facial biometrics has already seen impressive results; however, studying these techniques is necessary to determine how effective these methods are within various parameters. A Convolutional Neural Network (CNN) is a powerful classification approach which is often used for image identification and verification. Quite recently, CNNs have shown great promise in the area of facial image recognition. The comparative study proposed in this paper offers an in-depth analysis of several state-of-the-art deep learning based- facial recognition technologies, to determine via accuracy and other metrics which of those are most effective. In our study, VGG-16 and VGG-19 showed the highest levels of image recognition accuracy, Citation: Gwyn, T.; Roy, K.; Atay, M. as well as F1-Score.
    [Show full text]
  • Recurrent Neural Network
    Recurrent Neural Network TINGWU WANG, MACHINE LEARNING GROUP, UNIVERSITY OF TORONTO FOR CSC 2541, SPORT ANALYTICS Contents 1. Why do we need Recurrent Neural Network? 1. What Problems are Normal CNNs good at? 2. What are Sequence Tasks? 3. Ways to Deal with Sequence Labeling. 2. Math in a Vanilla Recurrent Neural Network 1. Vanilla Forward Pass 2. Vanilla Backward Pass 3. Vanilla Bidirectional Pass 4. Training of Vanilla RNN 5. Vanishing and exploding gradient problems 3. From Vanilla to LSTM 1. Definition 2. Forward Pass 3. Backward Pass 4. Miscellaneous 1. More than Language Model 2. GRU 5. Implementing RNN in Tensorflow Part One Why do we need Recurrent Neural Network? 1. What Problems are Normal CNNs good at? 2. What is Sequence Learning? 3. Ways to Deal with Sequence Labeling. 1. What Problems are CNNs normally good at? 1. Image classification as a naive example 1. Input: one image. 2. Output: the probability distribution of classes. 3. You need to provide one guess (output), and to do that you only need to look at one image (input). P(Cat|image) = 0.1 P(Panda|image) = 0.9 2. What is Sequence Learning? 1. Sequence learning is the study of machine learning algorithms designed for sequential data [1]. 2. Language model is one of the most interesting topics that use sequence labeling. 1. Language Translation 1. Understand the meaning of each word, and the relationship between words 2. Input: one sentence in German input = "Ich will stark Steuern senken" 3. Output: one sentence in English output = "I want to cut taxes bigly" (big league?) 2.
    [Show full text]
  • Neural Networks and Backpropagation
    10-601 Introduction to Machine Learning Machine Learning Department School of Computer Science Carnegie Mellon University Neural Networks and Backpropagation Neural Net Readings: Murphy -- Matt Gormley Bishop 5 Lecture 20 HTF 11 Mitchell 4 April 3, 2017 1 Reminders • Homework 6: Unsupervised Learning – Release: Wed, Mar. 22 – Due: Mon, Apr. 03 at 11:59pm • Homework 5 (Part II): Peer Review – Expectation: You Release: Wed, Mar. 29 should spend at most 1 – Due: Wed, Apr. 05 at 11:59pm hour on your reviews • Peer Tutoring 2 Neural Networks Outline • Logistic Regression (Recap) – Data, Model, Learning, Prediction • Neural Networks – A Recipe for Machine Learning Last Lecture – Visual Notation for Neural Networks – Example: Logistic Regression Output Surface – 2-Layer Neural Network – 3-Layer Neural Network • Neural Net Architectures – Objective Functions – Activation Functions • Backpropagation – Basic Chain Rule (of calculus) This Lecture – Chain Rule for Arbitrary Computation Graph – Backpropagation Algorithm – Module-based Automatic Differentiation (Autodiff) 3 DECISION BOUNDARY EXAMPLES 4 Example #1: Diagonal Band 5 Example #2: One Pocket 6 Example #3: Four Gaussians 7 Example #4: Two Pockets 8 Example #1: Diagonal Band 9 Example #1: Diagonal Band 10 Example #1: Diagonal Band Error in slides: “layers” should read “number of hidden units” All the neural networks in this section used 1 hidden layer. 11 Example #1: Diagonal Band 12 Example #1: Diagonal Band 13 Example #1: Diagonal Band 14 Example #1: Diagonal Band 15 Example #2: One Pocket
    [Show full text]
  • Deep Layer Aggregation
    Deep Layer Aggregation Fisher Yu Dequan Wang Evan Shelhamer Trevor Darrell UC Berkeley Abstract Visual recognition requires rich representations that span levels from low to high, scales from small to large, and resolutions from fine to coarse. Even with the depth of fea- Dense Connections Feature Pyramids tures in a convolutional network, a layer in isolation is not enough: compounding and aggregating these representa- tions improves inference of what and where. Architectural + efforts are exploring many dimensions for network back- bones, designing deeper or wider architectures, but how to best aggregate layers and blocks across a network deserves Deep Layer Aggregation further attention. Although skip connections have been in- Figure 1: Deep layer aggregation unifies semantic and spa- corporated to combine layers, these connections have been tial fusion to better capture what and where. Our aggregation “shallow” themselves, and only fuse by simple, one-step op- architectures encompass and extend densely connected net- erations. We augment standard architectures with deeper works and feature pyramid networks with hierarchical and aggregation to better fuse information across layers. Our iterative skip connections that deepen the representation and deep layer aggregation structures iteratively and hierarchi- refine resolution. cally merge the feature hierarchy to make networks with better accuracy and fewer parameters. Experiments across architectures and tasks show that deep layer aggregation improves recognition and resolution compared to existing riers, different blocks or modules have been incorporated branching and merging schemes. to balance and temper these quantities, such as bottlenecks for dimensionality reduction [29, 39, 17] or residual, gated, and concatenative connections for feature and gradient prop- agation [17, 38, 19].
    [Show full text]
  • Chapter 10: Artificial Neural Networks
    Chapter 10: Artificial Neural Networks Dr. Xudong Liu Assistant Professor School of Computing University of North Florida Monday, 9/30/2019 1 / 17 Overview 1 Artificial neuron: linear threshold unit (LTU) 2 Perceptron 3 Multi-Layer Perceptron (MLP), Deep Neural Networks (DNN) 4 Learning ANN's: Error Backpropagation Overview 2 / 17 Differential Calculus In this course, I shall stick to Leibniz's notation. The derivative of a function at an input point, when it exists, is the slope of the tangent line to the graph of the function. 2 0 dy Let y = f (x) = x + 2x + 1. Then, f (x) = dx = 2x + 2. A partial derivative of a function of several variables is its derivative with respect to one of those variables, with the others held constant. 2 2 @z Let z = f (x; y) = x + xy + y . Then, @x = 2x + y. The chain rule is a formula for computing the derivative of the composition of two or more functions. dz dz dy 0 0 Let z = f (y) and y = g(x). Then, dx = dy · dx = f (y) · g (x). 1 Nice thing about the sigmoid function f (x) = 1+e−x : f 0(x) = f (x) · (1 − f (x)). Some Preliminaries 3 / 17 Artificial Neuron Artificial neuron, also called linear threshold unit (LTU), by McCulloch and Pitts, 1943: with one or more numeric inputs, it produces a weighted sum of them, applies an activation function, and outputs the result. Common activation functions: step function and sigmoid function. Artificial Neuron 4 / 17 Linear Threshold Unit (LTU) Below is an LTU with the activation function being the step function.
    [Show full text]
  • Increasing Neurons Or Deepening Layers in Forecasting Maximum Temperature Time Series?
    atmosphere Article Increasing Neurons or Deepening Layers in Forecasting Maximum Temperature Time Series? Trang Thi Kieu Tran 1, Taesam Lee 1,* and Jong-Suk Kim 2,* 1 Department of Civil Engineering, ERI, Gyeongsang National University, 501 Jinju-daero, Jinju 660-701, Korea; [email protected] 2 State Key Laboratory of Water Resources and Hydropower Engineering Science, Wuhan University, Wuhan 430072, China * Correspondence: [email protected] (T.L.); [email protected] (J.-S.K.) Received: 8 September 2020; Accepted: 7 October 2020; Published: 9 October 2020 Abstract: Weather forecasting, especially that of extreme climatic events, has gained considerable attention among researchers due to their impacts on natural ecosystems and human life. The applicability of artificial neural networks (ANNs) in non-linear process forecasting has significantly contributed to hydro-climatology. The efficiency of neural network functions depends on the network structure and parameters. This study proposed a new approach to forecasting a one-day-ahead maximum temperature time series for South Korea to discuss the relationship between network specifications and performance by employing various scenarios for the number of parameters and hidden layers in the ANN model. Specifically, a different number of trainable parameters (i.e., the total number of weights and bias) and distinctive numbers of hidden layers were compared for system-performance effects. If the parameter sizes were too large, the root mean square error (RMSE) would be generally increased, and the model’s ability was impaired. Besides, too many hidden layers would reduce the system prediction if the number of parameters was high. The number of parameters and hidden layers affected the performance of ANN models for time series forecasting competitively.
    [Show full text]
  • Introduction to Machine Learning CMU-10701 Deep Learning
    Introduction to Machine Learning CMU-10701 Deep Learning Barnabás Póczos & Aarti Singh Credits Many of the pictures, results, and other materials are taken from: Ruslan Salakhutdinov Joshua Bengio Geoffrey Hinton Yann LeCun 2 Contents Definition and Motivation History of Deep architectures Deep architectures Convolutional networks Deep Belief networks Applications 3 Deep architectures Defintion: Deep architectures are composed of multiple levels of non-linear operations, such as neural nets with many hidden layers. Output layer Hidden layers Input layer 4 Goal of Deep architectures Goal: Deep learning methods aim at . learning feature hierarchies . where features from higher levels of the hierarchy are formed by lower level features. edges, local shapes, object parts Low level representation 5 Figure is from Yoshua Bengio Neurobiological Motivation Most current learning algorithms are shallow architectures (1-3 levels) (SVM, kNN, MoG, KDE, Parzen Kernel regression, PCA, Perceptron,…) The mammal brain is organized in a deep architecture (Serre, Kreiman, Kouh, Cadieu, Knoblich, & Poggio, 2007) (E.g. visual system has 5 to 10 levels) 6 Deep Learning History Inspired by the architectural depth of the brain, researchers wanted for decades to train deep multi-layer neural networks. No successful attempts were reported before 2006 … Researchers reported positive experimental results with typically two or three levels (i.e. one or two hidden layers), but training deeper networks consistently yielded poorer results. Exception: convolutional neural networks, LeCun 1998 SVM: Vapnik and his co-workers developed the Support Vector Machine (1993). It is a shallow architecture. Digression: In the 1990’s, many researchers abandoned neural networks with multiple adaptive hidden layers because SVMs worked better, and there was no successful attempts to train deep networks.
    [Show full text]
  • Introduction to Machine Learning Deep Learning
    Introduction to Machine Learning Deep Learning Barnabás Póczos Credits Many of the pictures, results, and other materials are taken from: Ruslan Salakhutdinov Joshua Bengio Geoffrey Hinton Yann LeCun 2 Contents Definition and Motivation Deep architectures Convolutional networks Applications 3 Deep architectures Defintion: Deep architectures are composed of multiple levels of non-linear operations, such as neural nets with many hidden layers. Output layer Hidden layers Input layer 4 Goal of Deep architectures Goal: Deep learning methods aim at ▪ learning feature hierarchies ▪ where features from higher levels of the hierarchy are formed by lower level features. edges, local shapes, object parts Low level representation 5 Figure is from Yoshua Bengio Theoretical Advantages of Deep Architectures Some complicated functions cannot be efficiently represented (in terms of number of tunable elements) by architectures that are too shallow. Deep architectures might be able to represent some functions otherwise not efficiently representable. More formally: Functions that can be compactly represented by a depth k architecture might require an exponential number of computational elements to be represented by a depth k − 1 architecture The consequences are ▪ Computational: We don’t need exponentially many elements in the layers ▪ Statistical: poor generalization may be expected when using an insufficiently deep architecture for representing some functions. 9 Theoretical Advantages of Deep Architectures The Polynomial circuit: 10 Deep Convolutional Networks 11 Deep Convolutional Networks Compared to standard feedforward neural networks with similarly-sized layers, ▪ CNNs have much fewer connections and parameters ▪ and so they are easier to train, ▪ while their theoretically-best performance is likely to be only slightly worse.
    [Show full text]